Stockfish Testing Queue

Finished - 1741 tests

15-10-28 SC opposite_bishops_simple diff
LLR: -2.95 (-2.94,2.94) [-3.00,1.00]
Total: 29414 W: 5476 L: 5686 D: 18252
sprt @ 10+0.1 th 1 Variation on Stephane's idea of removing ad-hoc logic for opposite bishops. Take 2.
15-10-28 SC opposite_bishops_simple diff
LLR: -2.95 (-2.94,2.94) [-3.00,1.00]
Total: 25206 W: 4700 L: 4903 D: 15603
sprt @ 10+0.1 th 1 Variation on Stephane's idea of removing ad-hoc logic for opposite bishops.
15-10-29 SC quadratic_initiative_tu diff
4002/50000 iterations
8403/100000 games played
100000 @ 10+0.1 th 1 Try to tune a generalization of evaluate_initiative: every contribution is now a score and also quadratic terms are considered.
15-10-29 SC opposite_bishops_simple diff
LLR: -2.96 (-2.94,2.94) [-3.00,1.00]
Total: 7503 W: 1341 L: 1513 D: 4649
sprt @ 10+0.1 th 1 A last go with tuned values.
15-10-28 SC oppbis_simple_tuning diff
58047/50000 iterations
100000/100000 games played
100000 @ 5+0.1 th 1 Variation on Stephane's idea of removing ad-hoc logic for opposite bishops. Try to get the right coefficients by tuning at fast time control.
15-10-28 SC moreTuned diff
LLR: -2.95 (-2.94,2.94) [0.00,4.00]
Total: 54730 W: 10623 L: 10585 D: 33522
sprt @ 10+0.1 th 1 Results from SPSA tuning session. Keep symmetrized pawn values and remove history tuned values, according to mixed suggestions by VOne and Rocky. Take 4.
15-10-27 SC moreTuned diff
LLR: -2.94 (-2.94,2.94) [0.00,4.00]
Total: 14915 W: 2827 L: 2932 D: 9156
sprt @ 10+0.1 th 1 Results from SPSA tuning session. Restore original and isolated pawn values. Take 3.
15-10-27 SC moreTuned diff
LLR: -2.96 (-2.94,2.94) [0.00,4.00]
Total: 21513 W: 4160 L: 4242 D: 13111
sprt @ 10+0.1 th 1 Results from SPSA tuning session. Remove asymmetries, take 2.
15-10-27 SC moreTuned diff
LLR: -2.95 (-2.94,2.94) [0.00,4.00]
Total: 7438 W: 1381 L: 1513 D: 4544
sprt @ 10+0.1 th 1 Results from SPSA tuning session. Asymmetries in the pawn evaluation are interesting, as well new weights for cmh vs h in move ordering.
15-10-20 SC moreTuning diff
60500/150000 iterations
123740/300000 games played
300000 @ 40+0.4 th 1 As I was twice lucky with tuning, let me give a third go with other rarely considered parameters. Framework is mostly idle and a lot of machines want to work. Low throughput, so it uses only time when fishtest is idle.
15-10-26 SC log_formula_general diff
ELO: -214.05 +-103.1 (95%) LOS: 0.0%
Total: 31 W: 1 L: 18 D: 12
10000 @ 10+0.1 th 3 Try a threads independent generalization of the log formulas I have seen, based on my excel table https://docs.google.com/spreadsheets/d/1Ub3YsFdK_40Cp0DuoVhTVwbZnJXWfXgFwmxb5CN0gK4/edit#gid=0 Check whether it is a obvious regression on 3 threads (which is the thread count for which the change is largest).
15-10-18 SC assorted_tuning diff
LLR: 2.95 (-2.94,2.94) [0.00,4.00]
Total: 21577 W: 3507 L: 3289 D: 14781
sprt @ 40+0.4 th 1 Retest assorted tuning without time management as discussed in https://github.com/official-stockfish/Stockfish/pull/464 LTC
15-10-18 SC assorted_tuning diff
LLR: 3.07 (-2.94,2.94) [0.00,4.00]
Total: 15124 W: 2974 L: 2756 D: 9394
sprt @ 10+0.1 th 1 Retest assorted tuning without time management as discussed in https://github.com/official-stockfish/Stockfish/pull/464
15-10-17 SC initiative_pawns diff
LLR: -2.97 (-2.94,2.94) [0.00,5.00]
Total: 18987 W: 3604 L: 3647 D: 11736
sprt @ 10+0.1 th 1 Try the opposite.
15-10-17 SC initiative_pawns diff
LLR: -2.95 (-2.94,2.94) [0.00,5.00]
Total: 8467 W: 1596 L: 1685 D: 5186
sprt @ 10+0.1 th 1 Only use pawn count from strong side in scaling endgame.
15-10-16 SC assorted_tuning diff
LLR: 3.01 (-2.94,2.94) [0.00,5.00]
Total: 13886 W: 2239 L: 2063 D: 9584
sprt @ 40+0.4 th 1 There was some almost passed tuning attempts in the last month. Collect them and give them a second chance. LTC.
15-10-15 SC assorted_tuning diff
LLR: 2.96 (-2.94,2.94) [0.00,5.00]
Total: 8800 W: 1814 L: 1646 D: 5340
sprt @ 10+0.1 th 1 There was some almost passed tuning attempts in the last month. Collect them and give them a second chance.
15-10-04 SC scale_factor_tunable diff
LLR: 2.96 (-2.94,2.94) [0.00,4.00]
Total: 36089 W: 5589 L: 5331 D: 25169
sprt @ 60+0.05 th 1 Values after 183k iterations. Let us see. LTC.
15-10-03 SC scale_factor_tunable diff
LLR: 2.96 (-2.94,2.94) [0.00,4.00]
Total: 45401 W: 8590 L: 8274 D: 28537
sprt @ 15+0.05 th 1 Values after 183k iterations. Let us see.
15-09-26 SC scale_factor_tuning diff
89027/100000 iterations
184197/200000 games played
200000 @ 60+0.05 th 1 As we have a lot of machines active, let me submit a very long LTC tuning session on rarely considered parameters, just in case we left some ELO lying around. Low throughput, such that it kicks in only if no Priority 0 stuff is waiting. Maybe we can get an answer before next TCEC stage.
15-09-26 SC multipv_search diff
LLR: -3.23 (-2.94,2.94) [0.00,5.00]
Total: 12112 W: 2215 L: 2300 D: 7597
sprt @ 15+0.05 th 1 Last attempt.
15-09-26 SC multipv_search diff
LLR: -2.95 (-2.94,2.94) [0.00,5.00]
Total: 2951 W: 463 L: 575 D: 1913
sprt @ 15+0.05 th 1 I realized that I have messed up with the logic. Take 3, also taking into account BestMoveChanges.
15-09-26 SC multipv_search diff
LLR: -2.95 (-2.94,2.94) [0.00,5.00]
Total: 4113 W: 720 L: 828 D: 2565
sprt @ 15+0.05 th 1 Search with multipv at low depth. When we have plenty of time available, let us search in multipv. Maybe it help sorting moves. Scale multipv proportionally to sqrt of remaining time. Use more aggressive values. Take 2.
15-09-25 SC multipv_search diff
LLR: -2.96 (-2.94,2.94) [0.00,5.00]
Total: 15840 W: 2849 L: 2907 D: 10084
sprt @ 15+0.05 th 1 Search with multipv at low depth When we have plenty of time available, let us search in multipv. Maybe it help sorting moves. Scale multipv proportionally to sqrt of remaining time. Furiousfish ignored me for some reason.
15-09-24 SC imbalance_gp2 diff
LLR: -4.05 (-2.94,2.94) [0.00,5.00]
Total: 37303 W: 6876 L: 6887 D: 23540
sprt @ 15+0.05 th 1 Faster and tuned version.
15-09-21 SC imbalance_gp_tuning diff
19495/20000 iterations
38394/40000 games played
40000 @ 60+0.05 th 1 Tuning imbalance_gp after including ideas from Stephane. Since it failed LTC, I tune LTC at throughput 500. As suggested by Rocky640 reschedule with different settings (and less games).
15-09-20 SC imbalance_gp_tuning diff
11969/30000 iterations
23767/60000 games played
60000 @ 60+0.05 th 1 Tuning imbalance_gp after including ideas from Stephane. Since it failed LTC, I tune LTC at throughput 500. Bugfix (forgot to commit before pushing).
15-09-20 SC oppBishops_passedPawns diff
LLR: -2.95 (-2.94,2.94) [-3.00,1.00]
Total: 14027 W: 2472 L: 2653 D: 8902
sprt @ 15+0.05 th 1 I was wondering: we are making a lot scaling stuff with opposite color bishops, but the very important thing about them is that you cannot exploit passed pawns properly. So try to go and change evaluation of passed pawns using information about opposite bishops and remove all the scaling stuff.
15-09-19 SC imbalance_simp diff
LLR: -2.95 (-2.94,2.94) [-3.00,1.00]
Total: 7080 W: 1217 L: 1386 D: 4477
sprt @ 15+0.05 th 1 Approvers please read https://groups.google.com/forum/?fromgroups=#!topic/fishcooking/EIe-176l1hA before (if at all) approving
15-09-18 SC imbalance_gp diff
LLR: -2.94 (-2.94,2.94) [0.00,5.00]
Total: 25259 W: 3821 L: 3848 D: 17590
sprt @ 60+0.05 th 1 Give a 10% malus for material imbalance in middlegame and a 10% bonus in endgame. LTC.
15-09-18 SC imbalance_gp diff
LLR: -1.99 (-2.94,2.94) [0.00,5.00]
Total: 20636 W: 3761 L: 3757 D: 13118
sprt @ 15+0.05 th 1 Give a 10% boost for material imbalance in middlegame and a 10% malus in endgame
15-09-18 SC imbalance_gp diff
LLR: 2.95 (-2.94,2.94) [0.00,5.00]
Total: 16030 W: 3056 L: 2861 D: 10113
sprt @ 15+0.05 th 1 Give a 10% malus for material imbalance in middlegame and a 10% bonus in endgame
15-09-17 SC capture_gp diff
LLR: -3.00 (-2.94,2.94) [0.00,5.00]
Total: 9665 W: 1768 L: 1854 D: 6043
sprt @ 15+0.05 th 1 Attempt on pawn count based capture ordering with quadratic dependency on pawn count. Gave very good results locally on TTD measures.
15-09-15 SC phalanx_tuned diff
LLR: -2.95 (-2.94,2.94) [0.00,5.00]
Total: 14062 W: 2663 L: 2727 D: 8672
sprt @ 15+0.05 th 1 Why we have phalanx bonus for phalanxes on RANK_2?
15-09-15 SC phalanx_1 diff
LLR: -2.96 (-2.94,2.94) [0.00,5.00]
Total: 5661 W: 1047 L: 1149 D: 3465
sprt @ 15+0.05 th 1 A furthter attempt on phalanxes with file bonuses.
15-09-13 SC phalanx_1 diff
LLR: -2.95 (-2.94,2.94) [0.00,5.00]
Total: 33730 W: 6271 L: 6250 D: 21209
sprt @ 15+0.05 th 1 Values from SPSA session
15-09-13 SC capture_preinit diff
LLR: -2.95 (-2.94,2.94) [0.00,5.00]
Total: 5293 W: 961 L: 1064 D: 3268
sprt @ 15+0.05 th 1 I want to follow some suggestions by VoyagerOne in https://groups.google.com/forum/?fromgroups=#!topic/fishcooking/IR6yc1KqcwM and I start by giving a second try to rankPenalty = 100 + 16*pawnCount, but with preinitialized scores. If this minor speed up also fails, I can go for more complicated formulas.
15-09-13 SC phalanx_1 diff
LLR: -2.97 (-2.94,2.94) [0.00,5.00]
Total: 13885 W: 2574 L: 2640 D: 8671
sprt @ 15+0.05 th 1 A variant of file bonuses for phalanx: use much higher values, but only for RANK_4 or more.
15-09-12 SC phalanx_tuned diff
LLR: -2.95 (-2.94,2.94) [0.00,4.00]
Total: 33864 W: 6317 L: 6356 D: 21191
sprt @ 15+0.05 th 1 Values from phalanx tuning session without file bonuses
15-09-09 SC phalanx_tune diff
49192/50000 iterations
99341/100000 games played
100000 @ 30+0.05 th 1 In June there was this phalanx_1 test passing STC in less than 4000 games, but failing LTC. Tune it at intermediate TC. http://tests.stockfishchess.org/tests/view/557345b10ebc5940ca5d7063
15-09-11 SC alekhine_gun diff
LLR: -2.95 (-2.94,2.94) [0.00,5.00]
Total: 15806 W: 2928 L: 2985 D: 9893
sprt @ 15+0.05 th 1 Try to improve queen_battery patch with the Alekhine's gun. See e.g https://en.wikipedia.org/wiki/Alekhine's_gun
15-09-08 SC capture_gp diff
LLR: -2.96 (-2.94,2.94) [0.00,5.00]
Total: 33756 W: 6311 L: 6290 D: 21155
sprt @ 15+0.05 th 1 I just noticed a bug in my code for take 7. This should be a fixed version.
15-09-07 SC capture_gp diff
LLR: -3.13 (-2.94,2.94) [0.00,5.00]
Total: 53803 W: 10002 L: 9902 D: 33899
sprt @ 15+0.05 th 1 Last try on game phase based rank penalty in capture ordering. 150 + 10*opponentPawns (local tuning by SEE approximation, like MVV + rank was tuned back then). But take 200 for RANK_8 as suggested by linear regression of SEE. Take 7 and after that I leave it alone.
15-09-07 SC capture_gp diff
LLR: -2.95 (-2.94,2.94) [0.00,5.00]
Total: 37925 W: 7108 L: 7068 D: 23749
sprt @ 15+0.05 th 1 Game phase based rank penalty in capture ordering. 150 + 10*opponentPawns (local tuning by SEE approximation, like MVV + rank was tuned back then). Take 5.
15-09-05 SC capture_gp diff
LLR: -2.96 (-2.94,2.94) [0.00,5.00]
Total: 78925 W: 14877 L: 14658 D: 49390
sprt @ 15+0.05 th 1 Game phase based rank penalty in capture ordering. 100 + 16*opponentPawns (interpolation between base and current patch, which will probably fail yellow). Take 4.
15-09-07 SC capture_gp diff
LLR: -2.97 (-2.94,2.94) [0.00,5.00]
Total: 5659 W: 1012 L: 1114 D: 3533
sprt @ 15+0.05 th 1 Opponent pawn count based capture ordering. Try to boost it setting rank penalty to 0 in case of promotion move. Take 5. Now with the correct base bench.
15-09-06 SC capture_gp_tuning diff
27747/30000 iterations
55934/60000 games played
60000 @ 30+0.05 th 1 Looks like the idea of scaling down rank penalty with number of pawns of opposite side could make it. Tune it with SPSA. I will decrease priority of current test.
15-09-03 SC capture_gp diff
LLR: -2.95 (-2.94,2.94) [0.00,5.00]
Total: 54203 W: 10257 L: 10145 D: 33801
sprt @ 15+0.05 th 1 Ok, optimal rank penalty does not depend on aggressor piece. Maybe depends on game phase? Take 1: opponent pawn count.
15-09-03 SC capture_gp diff
LLR: -2.94 (-2.94,2.94) [0.00,5.00]
Total: 11617 W: 2135 L: 2210 D: 7272
sprt @ 15+0.05 th 1 Ok, optimal rank penalty does not depend on aggressor piece. Maybe depends on game phase? Take 3: opponent non pawn material.
15-09-03 SC capture_gp diff
LLR: -2.96 (-2.94,2.94) [0.00,5.00]
Total: 4877 W: 847 L: 952 D: 3078
sprt @ 15+0.05 th 1 Ok, optimal rank penalty does not depend on aggressor piece. Maybe depends on game phase? Take 1: opponent piece count.